Académique Documents
Professionnel Documents
Culture Documents
(1)
3.2. Geographical distance
Despite the wide diffusion of communication technologies which shrinks perceived distances, geographical
distances still play a role when it comes to the propensity to cooperate and to select a cooperation partner
(Leamer & Storper 2001). In various industries we find tendencies for an uneven distribution of firms in
7
space. This holds in particular for high-tech industries (Audretsch & Feldman 1996). In figure 1 the clustered
geographical dispersion of the German automotive industry is shown.
Figure 1: Geographical distribution of automotive firms in Germany
The geographical clustering of firms influence interaction patterns (e.g. Weterings & Boschma 2009,
Hoekman, Frenken & Van Oort 2009). Shorter distances provide more opportunities to meet which helps to
develop trust as a prerequisite for the willingness to exchange knowledge, in particular tacit knowledge
(Howells 2002). Face-to-face interaction facilitates learning processes and interactive learning. According to
Glckler (2007) there are two channels by which distance exerts influence: First, short distances positively
affect the formation of interfirm networks. Note, however, that it is not the geographical distance as such
which influences network formation. Instead, the possibilities and preferences of agents to communicate
matter (Storper & Venables 2004). Also, the infrastructure and possibilities to travel faster are to be taken
into account (Marquis 2003). Thus, there is a potential relation between geographical distances and the
possibilities and propensities to form fruitful agreements of interaction which, however, is by far not
compulsory. Second, locations may play a role by providing opportunities to access specific and locally
bounded resources (e.g. specialized workforce) and regional unequal distributed opportunities for
economic development (Sayer 1991, Bathelt & Glckler 2005).
It follows that firms which are located in relative spatial proximity have a higher propensity to cooperate
compared to other pairs (H2). In order to form a pairwise distance matrix, geographical distances between
8
all pairs of actors (dist
ij
) have been retrieved by a specific search routine from the web navigation service
Google Maps and logarithmized with the natural logarithm (Formula 2).
(2)
3.3. Absorptive capacity
Absorptive capacity reflects the ability to evaluate, assimilate and exploit knowledge from external sources
(Cohen & Levinthal 1990). If a firm has already accumulated knowledge in the same or related fields it is
easier for it to recognize, evaluate, assimilate and apply external knowledge. Thus learning is cumulative,
and learning performance is greatest when the object of learning is related to what is already known
(Cohen & Levinthal 1990, p. 131). For our case of testing the drivers of network dynamics two
considerations are relevant: first, firms do have advantages in learning new things in fields which are close
to their knowledge-base, while it is rather difficult in fields which are completely different. Second, the
characteristics of an actors knowledge-base changes only incrementally due to the fact that learning takes
only place in fields that are related and somewhat similar to already explored fields (Cohen & Levinthal
1990). Zahra and George (2002) stress that dynamic absorptive capacities allow firms to build up the
required knowledge to develop other organizational capabilities and to broaden their knowledge-base in
time. Consequently, we expect firms which have higher levels of absorptive capacity to also have a higher
propensity to collaborate as they face more opportunities to benefit from external knowledge (H3). The
absorptive capacity (v
absorpcai(t)
) is approximated by taking the natural logarithm of the number of patents
(NbPatents
i(t:t-5)
) a firm applied for in the five years prior to the observation point. Accordingly absorptive
capacities of actors increase with the accumulated patenting activity with diminishing rates (Formula 3).
) (3)
3.4. Technological proximity
Besides geographical proximity also technological proximity among actors matters for cooperation. Here
we refer to the notion of cooperation partner similarity (McPherson, Smith-Lovin & Cook 2001) in a
network context. According to this concept, similar nodes have a higher probability to form a tie. However,
similarity can refer to various dimensions. Partner could be similar with regard to technological, knowledge-
related, organisational or financial characteristics or comparable in terms of reputation and status. For
instance, Gulati (1995b) and Rothaermel and Boeker (2008) demonstrate that status similarity increases the
rate of tie formations in interorganizational networks. As we focus on innovation networks the similarity of
the technological knowledge-base is of outmost importance. The notion of technological proximity refers to
shared technological experiences and knowledge-bases (Knoben & Oerlemans 2006). Thus, it does not
express the similarity of technological equipment, processes etc., but reflects the similarity of the
2
For all dyadic covariates
9
underlying knowledge-bases. This understanding is somewhat similar to the concept of cognitive proximity
as described by Boschma (2005) even though cognitive proximity is more comprehensive.
Knowledge-base similarity facilitates learning. In addition, it sharpens the senses for the perception of
emerging technological trends (Zeller 2004). From Cohen and Levinthal (1990) it can be inferred that
effective learning of an organization necessitates a certain degree of similar problem perception and
assimilation of new knowledge but at the same time some degree of diversity is useful to develop new
ideas based on the acquired knowledge. This idea transferred to the dyadic level suggests hypothesis H4a
that cooperating firms must, for effective learning, have similar knowledge-bases which reflect a common
understanding of problems and increases the capacity to absorb each others knowledge (Colombo 2003).
On the other hand, invention and innovation can be understood as a new combination of existing
knowledge which would require the combination of more different knowledge-bases which is our
hypothesis H4b. Not surprisingly, Nooteboom et al. (2007) find a U-shaped curve for an optimal cognitive
distance which is, as mentioned earlier, conceptually close to the technological distance.
For the calculation of distances of firms in technological space, we apply the Euclidean distance (E) measure
based on a firms patent portfolio which encompasses all EPO patents filed not more than 5 year prior to
the observation point. In a first step, a vector is calculated which puts each firm in an N-dimensional vector
space. The number of dimensions N results from the number of 3-digit IPC classes in which all firms filed
patents (priority filling). The firm vector p is given by the relative share of patents a firms has in the N
patent classes. For instance, if N is only two (e.g. B60 and B29) and a firm has 40% of its patents in class B60
and 60% in B29 the vector is (p
B60
; p
B29
) = (0,4; 0,6). In a second step, differences between vectors
representing distances in the technology space are calculated. Thus the technological distance (w
techdisij
)
between firms i and j is calculated as formula (4) suggests:
(4)
3.5. Experience with cooperation
A further factor we examine is the firms experience with cooperation. We assume for Hypothesis H5 that a
large record of collaborative activities signals a larger attractiveness for further collaboration. This reflects
that from outside it is rather difficult to scan a firms valuable resources, in particular the knowledge. Thus,
a firm which has been often involved in cooperative projects signals to be a valuable partner with a good
reputation and established routines of collaboration. Cooperation capabilities are specific and not
transferable resources which can enhance a firms ability to identify partner, initiate collaborations and
manage successfully the partnerships (e.g. Makadok 2001). Experienced firms have implemented
collaboration management functions whose task is to coordinate the portfolio of different types of alliances
(Kale, Dyer & Singh 2002). Developing experience is time consuming because firms are forced to adapt
internal routines (Powell, Koput & Smith-Doerr 1996). However, it is worth the effort as it not only enables
10
a firm to become effectively embedded in a formal innovation network but also paves the ground for
likewise important informal collaboration (Pyka 2000). In formula (5) we measure the experience of a firm
(v
expi(t)
) with the frequency of participation in the subsidised R&D projects with partners (NbR&D
projectsi(t:1998)
)
from within or outside the automotive sample in the years 1998 to 2007.
(5)
3.6. Knowledge-base modularity
Finally, modularity constitutes a basic evolutionary principle (Pyka 2002). It has been mostly studied with
respect to product architecture and organizational structures (see for instance Sanchez & Mahoney 1996,
Baldwin & Clark 2000, Schilling 2000, Ethiraj & Levinthal 2004). Modularity effects concerning the
knowledge structure, however, so far have been of minor interest. Some studies identify a relation
between the structure of the organizational knowledge-base and innovation related outcomes. Ahuja and
Katila (2001), for instance, show that the size of the knowledge base is positively correlated with innovative
productivity. Lane and Lubatkin (1998) find that the degree to which two knowledge-bases overlap
influences positively the ability of mutual learning in cooperation.
With respect to the modularity of knowledge-bases, we suggest that not only the size or the relatedness
matter for the propensity to cooperate but also that the decomposability of the knowledge-bases into
modular knowledge substructures. The feature of modularity of a firms knowledge-base is approximated
by its degree of clustering and decomposability. The core argument is that firms not only try to find a
partner from which they can learn or which has a similar technological understanding, but they attempt to
link technologies. This combination of technologies is way easier if the technologies have a modular
structure. A decomposable knowledge-base enables researchers to conduct recombinant search processes
without getting trapped in complexity and endless combinatorial possibilities (Yayavaram & Ahuja 2008).
The recombinatorial possibilities become rapidly very large even with rather modest sized knowledge-
bases. Firms that search for an appropriate cooperation partner to combine elements of their own
knowledge-base with elements or a partners knowledge-base are confronted with a high level of
complexity, an overload of possibilities and uncertainty at the same time. Thus, modularity reduces time
and costly search process as compatible technologies can be identified more easily. From these
considerations we propose our hypothesis H6: the propensity of two firms to cooperate rises with the
possibility to structure their knowledge-base in a modular way.
To analyze the decomposability of a firms knowledge-base we consider it as a network of knowledge
elements (IPC classes) that are linked by patents (affiliation or co-occurrence network). We further assume
that a link emerges between technology classes once they are mentioned on the same patent (see Saviotti
2009 and Yayavaram & Ahuja 2008 for details). Yayavaram (2008, p. 334) states that the set of couplings
or ties together with the strength of the ties constitute the structure of a firms knowledge base. The
11
connection of knowledge elements is not only dichotomous but varies in its intensity. For instance, two
knowledge elements (IPC classes) A and B may appear ten times together on patents of a firm while the
elements A and C may only appear together once. Consequently, for our network representing the
knowledge-base of firms we use the frequency of co-occurrences as weights for the ties. The degree of
decomposability is reflected by a continuum of structures. From a non-modular to a highly modular
knowledge-base the links between knowledge elements become more equally distributed (Figure 2).
Figure 2: Relation between modularity and clustering
A high modular knowledge-base is characterized by some knowledge elements forming a dense cluster and
different clusters being independent (Figure 2 top right). Nearly decomposable structures (Simon 1962)
show some elements forming groups through dense links and some links between these clustered
knowledge elements (Figure 2 top middle). Baldwin & Clark (2000), Schilling (2000) and Weick (1976)
suggest that nearly decomposable structures benefit from a high potential for recombination, persistence
and adaptability. Yayavaram and Ahuja (2008) and March (1991) highlight that nearly decomposability
allows for a balanced search between depth and breadth. Finally, a non decomposable pattern does not
show identifiable clusters but the links seem to be arbitrarily distributed (Figure 2 top left).
The coupling of knowledge elements is caused by three distinct motives:
(i) there is a natural interdependence between some knowledge elements which mutually influence each
other;
(ii) search routines of firms may be directed to the coupling of certain knowledge elements while other
elements are used more independently;
12
(iii) finally, innovation processes are often recombinant, i.e. coupling different so far unrelated knowledge
elements.
As a measure for the level of modularity of a knowledge-base we calculate a slightly modified clustering
coefficient. The first step to calculate the modularity of the knowledge-bases is to construct a knowledge
network structure. For this purpose, we take IPC sub-classes (4-digit level) as nodes and add a tie between
nodes whenever the IPC sub-classes co-occurred in a patent. Doing this, we reconstruct the knowledge
network from patents for each of the analyzed 153 firms in two time windows encompassing five years
(1998-2002; 2002-2006). By neglecting the tie strengths (dichotomization of the adjacency matrix), we then
calculate the clustering coefficient (cc) for each node of a firms knowledge-base network.
The clustering coefficient for node (IPC sub-class) i with k
i
ties (CC
i
) is defined in formula 6:
(6)
The calculation includes n
i
, the number of ties between the k
i
neighbours of node i. The denominator
represents the maximum number of ties which are possible between the k
i
neighbours of node i.
In order to weight the IPC sub-classes which appear more often in the patent portfolio we calculate in (7)
the share (RSC
i
) of an IPC sub-class (C
i
) relative to all IPC sub-classes in the portfolio such as:
(7)
In a final step, the clustering indicator (CI
i
) for each firms (j) knowledge base is calculated in (8) by
multiplying the clustering coefficients CC
i
with the relative shares of the IPC sub-classes (RSC
I
) and summing
up the weighted clustering coefficients:
(8)
3.7. Control variables
Two important controls have been added to the model, both referring to a capacity effect. Larger firms can
coordinate more cooperation partners at the same time than smaller firms. Accordingly, the model needs
to control for firm size. For measuring the size, three categories have been created, namely large firms,
medium sized firms and small firms. Threshold levels are applied for the number of employees and/or the
annual turnover for the years 2002-2010. Data are taken from the companies websites and
Handelsregister excerpts (accessed via LexisNexis). For the categorization the usual classification is
chosen:
13
Category 1 (Large): > 249 employees; turnover 50 Mio.
Category 2 (Medium): 50-249 employees; turnover < 50 Mio.
Category 3 (Small): 10-49 employees; turnover < 10 Mio.
The second control variable is experience of firms in the industry. Likewise, older firms are more
experienced and can manage a higher number of collaborative projects. To approximate experience we
apply the natural logarithm of firm age.
4. Data sources and descriptive analysis
For empirical research of network evolution, the first challenge is to select the firms which are (potentially)
part of the network. This opens the discussion about the boundaries of the network (e.g. Laumann,
Marsden & Prensky (1983)). We aim for studying publicly funded research networks in the German
automotive industry. While it is relatively easy to filter German firms by their location (address), the
approach for covering an industry is more contentious. Since our reasoning is led by a knowledge-based
view of the firm, we start to build the sample based on a firms patent portfolio instead of applying a
standard industry classification like NACE. A scan of the patent portfolios (OECD June 2010) Regpat
database (which is a supplemented extraction from Patstat) of the German Original Equipment
Manufacturers (OEMs) and the largest suppliers shows that the 3-digit IPC class B60 is dominant in the
industry. Thus, we pick all firms which filed at least one patent application in this class within the
observation period 1998 to 2007 and pick out those which were exclusively operating in the market for
commercial vehicles or car accessory kits. This way we exclude all firms which were not directly related to
the production of passenger cars. We also exclude firms which have not been involved in at least one of the
examined research projects. This sampling resulted in 153 firms belonging to the network sample.
For the simulation implementation, networks are observed at six consecutive points in time (2002-2007)
resulting in six adjacency matrices reflecting the state of the network at the observation points. It is
generally challenging to find sources about interfirm networks, in particular for longitudinal network
studies. In our case, the networks are constructed from the database of the German Frderkatalog (R&D
subsidies catalogue) which contains rich information about research projects supported by the federal
government. The database is publicly accessible via the website www.foerderkatalog.de. Only those firms
are eventually picked for the analysis which participated in the observation period 1998-2007 at least one
time in a funded project. In the model a tie emerges between any two actors i and j if they participated in
the same project. Despite the fact that the database contains rich information about subsidised collective
research projects, it has been hardly used to conduct network research thus far (Broekel and Graf 2010).
Information about firms participating in joint subsidised projects documents research activities at an earlier
stage compared to patent data. R&D subsidies have become a frequently used instrument of innovation
14
policy makers to spur collaborative research for a number of reasons. First, due to the sheer scale of some
projects they cannot be afforded by single firms. Second, knowledge transfer from public to private
organisations is fostered by the participation of universities and other public research institutes such as
Max Planck and Fraunhofer. The projects listed in the Frderkatalog are considered to contribute to
knowledge transfer (Broekel & Graf 2010). The participants have to sign agreements explicitly stipulating
that gained knowledge within the project will be freely shared among the participants. They even have to
grant free access to their know-how and IPRs within the scope of the project. Furthermore, they commit to
actively collaborate with the aim to find new solutions (BMBF 2008). That this has worked out well and is
not to be considered a lip-service only is empirically shown (Fornahl, Broekel & Boschma 2011).
To construct networks from the project data, the following information were retrieved: name of the
project, starting and end date, name of the receiving/executing organisation. In addition, we find
information about the grant, the location of the receiving/executing organisation and a classification
number which divides funded technologies into different classes like biotechnology, energy etc. The title of
the project is important to separate cooperative (Verbundprojekt or Verbundvorhaben) from non
cooperative projects in which single organisations are funded. However, the title is not in all cases a clear
indication for a joint project. The database at hand is complementary to other sources like patent data or
publication data, in particular when it comes to longitudinal network studies (Broekel & Graf 2010) these
other data sources provide valuable information.
A possible drawback of our network data is the political determination, i.e. networks are to some extent
designed by political decisions to support certain key technologies that are considered as relevant for the
improvement of the competitiveness of the national economy. Innovation networks generated by policy
instruments might differ from emerging networks without external stimulus and confine results (Schn &
Pyka, 2012). Because publicly funded networks dissolve per definition after the funding period, windfall
profits are likely and long lasting linkages for knowledge transfer and learning might not appear. In many
cases the self-organizing networks are characterized by small world properties (Watts & Strogatz 1998)
which do not appear frequently in networks created by policy. For instance, small world properties were
found by Uzzi and Spiro (2005) for a network of Broadway musical artists, by Newman (2001) for networks
of scientific coauthoring in seven different scientific disciplines, by Fleming, King and Juda (2007) for patent
collaboration networks, by Davis, Yoo and Baker (2003) for the network of US company directors and by
Pyka, Gilbert and Ahrweiler (2007) for innovation networks in the biopharmaceutical industries. Finding
small world attributes in our innovation networks in the automotive industries would weaken this objection
towards publicly funded networks.
Small world networks are characterized by two features: (i) a high level of local clustering and (ii) a short
average path length between network actors. To test for small world characteristics, we draw on the Watts
and Strogatz (1998) approach which compares the observed networks path length and clustering
15
coefficient with the respective properties of a random network with the same size and same number of
ties. In contrast to random networks, small world networks are characterized by low clustering and low
path lengths. To quantify the comparison the small world quotient (Q) is applied. It is defined as the ratio of
the (global) clustering coefficient (CC) divided by the ratio of the average path length (PL). Measuring the
path length makes only sense in networks where all actors have at least one tie. Therefore, the largest
component is extracted from the full network. If the small world quotient is greater than 1.0 then the
network can be characterized as small world network. Table 1 shows that Q is in fact for all observed
networks larger than 1.0.
Furthermore, the critic concerning the usage of data describing publicly funded networks includes the idea
that granting schemes preselect eligible firms. However, in our case the data covers research processes at a
very early stage, something which cannot be achieved with patent data representing successful outcomes
of the research processes. Broekel & Graf (2010) argue that in fields of limited knowledge appropriability,
i.e. potential high technological spillovers, there is a strong need for public subsidies in order to create
incentives to invest. Accordingly, technology fields which meet these criteria are better covered by
subsidy data than by patent data.
Table 1: Small World test
Network
observation
2002
Network
observation
2003
Network
observation
2004
Network
observation
2005
Network
observation
2006
Network
observation
2007
Nodes
(largest
component)
56 52 46 50 59 66
Ties 326 270 222 236 426 452
CC 0,74 0,64 0,61 0,63 0,72 0,73
PL 2,42 2,55 2,66 2,62 2,60 2,59
Random Random Random Random Random Random
CCr 0,11 0,10 0,11 0,1 0,13 0,11
PLr 2,45 2,53 2,52 2,60 2,25 2,36
CC / CCr 6,73 6,40 5,55 6,30 5,54 6,64
PL / PLr 0,99 1,01 1,05 1,01 1,16 1,10
Q 6,81 6,35 5,26 6,25 4,79 6,05
Figure 3 and table 2 show a strong increase in the number of established ties between the observation
years 2000-2004 and 2001-2005 and in particular between the years 2001-2005 and 2002-2006. This can to
some extent be explained by an increased number of subsidised research projects because this policy
instrument gained in importance over the years. The number of disrupted as well as the number of stable
ties is faltering over the observation period.
16
1998-2002 1999-2003
2000-2004 2001-2005
2002-2006 2003-2007
Figure 3: Evolution of the automotive R&D innovation network (2002-2007)
In order to receive proper simulation results it is necessary to have a certain amount of change between
two consecutive waves. The assumption here is that changes in the network take place in a gradual
stepwise way rather than by sudden shocks. To ensure gradual change in the network data the Jaccard
17
index (Table 2) has been calculated (10) (M
11
= Number of pertained links; M
10
= Number of interrupted
links; M
01
= Number of formed links).
(10)
Based on experience with the application of the stochastic actor-based model of network evolution, the
value of the Jaccard Index should ideally be higher than 0.3 (Snijders, Van de Bunt & Steglich 2010). This is
the case for all observation periods which makes the data a good basis for simulation.
Table 2: Link development 2002-2007
Observation 00 01 10 11 Jaccard Index
12 10846 14 43 123 0.683
23 11174 14 39 98 0.649
34 11149 65 56 55 0.312
45 11085 120 25 95 0.396
56 11050 60 47 168 0.611
The density of the network (Table 3) is overall relatively low. It is slightly diminishing from 2002 to 2004 and
then rising again to the final year 2007. Likewise, the average degree centrality which indicates the average
number of established cooperative relations is decreasing in the first half and increasing again in the second
half. This tendency is confirmed by the number of ties which have been formed in the network.
Table 3: Density measures 2002-2007
Observation 2002 2003 2004 2005 2006 2007
Density 0.015 0.012 0.010 0.011 0.019 0.020
Average Degree
Centrality
2.258 1.839 1.483 1.611 2.886 3.060
Number of ties 166 137 112 120 215 228
5. The stochastic actor-based model of network evolution
The stochastic actor-based model for network dynamics (Snijders 1996) allows for the analysis of
longitudinal data describing network development and goes beyond the widespread models of network
evolution like e.g. scale-free networks of Barabasi and Albert (1999). The econometric application of scale
free networks is limited because of two reasons: First, the evolution of scale-free networks assumes as an
explanatory variable only the uneven distribution of the degrees of the actors in the network, other factors
18
shaping network evolution are not considered. Second, for understanding network dynamics it is not
sufficient to focus exclusively on the emergence of the network ties. Network dynamics are shaped to the
same extent by the dissolving of network linkages. The termination of network ties is explicitly considered
in the stochastic actor-based model.
In the stochastic actor-based model network ties are not understood as events but as states which persist
for some time. In our case of publicly funded R&D networks this is a realistic assumption as the projects
typically run for at least three years. The stochastic actor-based model is Markovian in the sense that the
current state of the network determines its further evolution. It follows a myopic stochastic optimization
rule (Snijders 2005). Thus, past events have no direct influence on the future which can be seen as a
violation of a cornerstone of evolutionary thinking. However, history still matters, because independent
variables inherently reflect historic information. This is the case for the experience with cooperation,
technological distances (if we assume path dependences), absorptive capacity and transitivity which is
based on structures that emerged in the past.
The basic approach demands the simulation of a large number of artificial networks which results in a
sample distribution of networks. The simulated networks then are compared with the observed network.
This allows for the estimation of parameters and standard errors to test the hypotheses. In this approach
(entire) networks are random variables (X) and have a probability distribution which is complex, thus it
cannot be approximated e.g. by a normal distribution. The actually observed network (x) is assumed to be
drawn from the population of all possible networks that are simulated based on a model that includes the
effects which are tested as drivers of network evolution.
The basic denotation is in line with standard network analysis. Networks are represented by an n x n
adjacency matrix
for m = 1, ..., M.
for all i. An
additional matrix, the composition change matrix, accounts for changes in the sample of firms. It includes
information about firms that enter or leave the network within the observation period because they are
only founded after the start of the observation or because they were dissolved, for instance due to an
acquisition or exit.
It is important to highlight that changes in tie variables are the dependent variables in the model. Modelling
network evolution is only meaningful if we have at least two observations, thus M must take a value 2.
The time parameter t is continuous. For the estimation of parameters it is however assumed that we
observe the network at least at two discrete points in time which gives it the character of a panel analysis.
The algorithm applied in the model can be traced back to earlier publications of Holland and Leinhardt
(1977), Wasserman (1980) and Leenders (1995). Yet, these prior models are rather limited in the structural
effect that can be modelled. The assumption of continuous time is advantageous for modelling tie
19
dependencies with tie formations that mutually depends on one another. Imagine the following example:
At t=0, in a group of three firms no firm cooperates. At t=1 the three firms started to cooperate and thus
form a triangle. In a model with only discrete time the emergence of the triangle structure could not be
explained, it just happened. In contrast, a continuous time model allows for a step by step, or better tie by
tie emergence of the observed triangle structure, for instance due to a transitive closure mechanism.
An actor which gets the possibility to change a tie is randomly chosen. Only one tie in the network can be
changed per iteration of the algorithm (Snijders 2001, Holland & Leinhardt 1977). That is, the change
process is broken down into the smallest possible components, called mini steps. This means that actors do
not coordinate tie changes but follow each other tie by tie and react to the stepwise changed network
structure. This assumption is somewhat debatable with respect to innovation networks that are based
without doubt - on some kind of coordination or negotiation. In this sense the algorithm has to be
considered as purely technical, which, however, allows for disentangling the importance of the various
effects tested for in the model.
The change process consists of two stochastic sub-processes. First, the frequency actors get the
opportunity to change a tie (change opportunity process) may depend on the position (centrality) of an
actor and on other covariates like experience or it is represented by a constant probability function.
Second, the tie change process which is determined by a probability function is again influenced by the
network position and the covariates of the ego but also of the other actors in the network. This simulation
model has the same underlying principles as other agent- or actor-based models. However, as the model is
used for statistical inference, it has to fulfil special requirements. First, we must be able to estimate the
parameters from the data in order to get a sound fit. Second, parsimony is a prerequisite, i.e. there should
not be any other ne details in the model else than what can be estimated from the data (Snijders, Van de
Bunt & Steglich 2010).
As proposed by Snijders (2001) the parameters of the model are estimated based on simulations with one
exemption. The network state of the first observation is not simulated but used as initial structure from
which on the change to the second observation is simulated. In other word, we model the change between
two consecutive observation points but we do not model the first observation. Parameters of the objective
function (referring to the tested effects) are obtained by applying an iterative Markov Chain Monte Carlo
(MCMC) algorithm (based on the method of moments). The stochastic approximation algorithm simulates
the evolution of the network and estimates parameters that minimize the deviation between observed and
simulated networks. During the iterations, the initial parameters of the model are incrementally adjusted
with the aim to find the best fit between simulated and observed networks. The final value of the
parameter determines the goodness of fit of the model and the standards errors. The MCMC estimations
can be interpreted like the results of a logistic regression which means that it has be controlled for all
20
potentially relevant variables that could influence change processes in the network (Balland, De Vaan &
Boschma 2012).
The change of network ties is determined by the objective function which expresses how the firm perceives
the network and evaluates the different change options. The aim of each actor is to increase the value of
the objective function which is determined by the ego-network, i.e. its direct (or indirect) ties and the
covariates of the other actors which are part of the network. For the processes of changing or keeping ties,
we have to consider the probabilities which are in turn dependent on the evaluation of possible changes in
the network in terms of ties and covariates. The way the objective function is constructed represents the
rules we assume to be relevant from the view of an actor when it makes choices. For any possible state of
the network the objective function takes a certain value. For higher values the probability increases that
the actor opts for this possible network state. Formally the objective function (9) is a linear combination of
a variety of components which are called effects and are explained in section 3 for our case.
(9)
The value of the objective function
= 0,
corresponding effects play no role in network evolution; if